Graph-based tracing of filamentary structured objects with applications in neuronal and retinal images

• Main Author: De, Jaydeep
• Other Authors: Li Cheng
• Format: Theses and Dissertations
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Pattern recognition; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/65993
• Institution: Nanyang Technological University

Filamentary structured objects are everywhere in nature and man-made architectures. These are found in the network of neurons, blood vessels, roads and others, which can scale from a range of few microns (in case of neurons) to thousands of kilometers (in case of roads). Image-based analysis of these filamentary structured objects are of great interest as they offer important information about the structure and connectivity of such networks which can help us in applications such as drug screening for neurological disorders or computer-aided diagnosis of diabetic retinopathy. However, the task is challenging due to the bottleneck of filament crossover issue, which essentially hinders the use of existing systems in large-scale applications. The aim of the thesis is to build a generic graph-based framework for tracing the images of filamentary structured objects. The framework consists of two steps: preprocessing step consisting of segmentation, skeleton extraction, digraph representation and tracing step using graph-based methods. The main focus of this thesis is the second step. After the preprocessing step, the skeleton is represented as a directed graph structure where skeleton segments are represented as nodes in the graph and there is an edge between two nodes if the corresponding skeletons are touching each other. Now, the goal is to cluster the graph into disjoint trees where the number of clusters depends on the number of tree structures in the filament. This is achieved by assuming that at least one node is labeled in each cluster of the graph. Given the structure of the graph and one labeled node per cluster, the problem becomes that of assigning class labels to the rest of the unlabeled nodes. The problem is formulated as a label propagation problem on a weighted digraph using the Matrix Forest Theorem. The normalized conductance of a rooted spanning converging forest of a digraph is used as a similarity score between the labeled and unlabeled nodes. The label propagation problem can also be formulated as a random walk on an Absorbing Markov Chain (AMC). After converting the original graph into an AMC, the Fundamental Matrix is computed, which is the expected number of visits from one node to another before absorption. For tracing problem, this is another kind of similarity score between the labeled and unlabeled nodes. For both approaches, the class label for an unlabeled node is the same as that of a labeled node with largest similarity. If the weight matrix is formulated as parametric functions then the parameters can be determined by maximizing the sum of the logarithm of expected number of visits from unlabeled nodes to the labeled nodes. For the scenario where the class labels for only very few nodes are known, the parameter estimation problem is solved by Expectation Maximization (EM) algorithm. For comparison purpose, the tracing problem is also formulated as a Maximum A Posteriori (MAP) inference problem in Undirected Graphical Models. Empirical analysis is conducted for all the approaches using in-house dataset of 2D neuronal images (Downloadable at http://web.bii.a-star.edu.sg/~jaydeepd/tracing.htm ) and publicly available benchmark dataset of retinal images which proves the superiority of my approach compared to current state of art methods as well as widely used commercial software.